Updated model card

README.md

@@ -24,9 +24,9 @@ license: apache-2.0
   - [Other Known Limitations](#other-known-limitations)
 - [Additional Information](#additional-information)
   - [Dataset Curators](#dataset-curators)
+  - [Citation Information](#citation-information)
 <!--- 
   - [Licensing Information](#licensing-information)
-  - [Citation Information](#citation-information)
   - [Contributions](#contributions)
 --->
 
@@ -40,6 +40,8 @@ license: apache-2.0
 
 We assembled a large-scale pretraining corpus, Genecorpus-30M, comprised of ~30 million human single cell transcriptomes from a broad range of tissues from publicly available data. This corpus was used for pretraining [Geneformer](https://huggingface.co/ctheodoris/Geneformer), a pretrained transformer model that enables context-aware predictions in settings with limited data in network biology.
 
+See [our manuscript](https://www.nature.com/articles/s41586-023-06139-9) for details.
+
 ### Supported Tasks
 
 This corpus was used for pretraining [Geneformer](https://huggingface.co/ctheodoris/Geneformer) and is compatible with pretraining or fine-tuning Geneformer or similar models.
@@ -54,9 +56,9 @@ Homo sapiens
 
 Genecorpus-30M is provided as tokenized data in the Huggingface Datasets structure, which is based on the Apache Arrow format. Each example within the dataset is composed of the rank value encoding for a single cell within the corpus. Rank value encodings provide a nonparametric representation of each single cell’s transcriptome, ranking genes by their expression within that cell normalized by their expression across the entire Genecorpus-30M. This method takes advantage of the many observations of each gene’s expression across Genecorpus-30M to prioritize genes that distinguish cell state. Specifically, this method will deprioritize ubiquitously highly-expressed housekeeping genes by normalizing them to a lower rank. Conversely, genes such as transcription factors that may be lowly expressed when they are expressed but highly distinguish cell state will move to a higher rank within the encoding. Furthermore, this rank-based approach may be more robust against technical artifacts that may systematically bias the absolute transcript counts value while the overall relative ranking of genes within each cell remains more stable.
 
-To accomplish this, we first calculated the nonzero median value of expression of each detected gene across all cells from the entire Genecorpus-30M. We aggregated the transcript count distribution for each gene, normalizing the gene transcript counts in each cell by the total transcript count of that cell to account for varying sequencing depth. We then normalized the genes in each single cell transcriptome by that gene’s non-zero median value of expression across Genecorpus-30M and ordered the genes by the rank of their normalized expression in that specific cell. Of note, we opted to use the nonzero median value of expression rather than include zeros in the distribution so as not to weight the value by tissue representation within Genecorpus-30M, assuming that a representative range of transcript values would be observed within the cells in which each gene was detected.
+To accomplish this, we first calculated the nonzero median value of expression of each detected gene across all cells from the entire Genecorpus-30M. We aggregated the transcript count distribution for each gene, normalizing the gene transcript counts in each cell by the total transcript count of that cell to account for varying sequencing depth. We then normalized the genes in each single cell transcriptome by that gene’s nonzero median value of expression across Genecorpus-30M and ordered the genes by the rank of their normalized expression in that specific cell. Of note, we opted to use the nonzero median value of expression rather than include zeros in the distribution so as not to weight the value by tissue representation within Genecorpus-30M, assuming that a representative range of transcript values would be observed within the cells in which each gene was detected.
 
-The rank value encodings for each single cell transcriptome were then tokenized based on a total vocabulary of 25,424 protein-coding or miRNA genes detected within Geneformer-30M. The token dictionary mapping each token ID to special tokens (<pad> and <mask>) or Ensembl IDs for each gene is included within the repository as a pickle file (token_dictionary.pickle).
+The rank value encodings for each single cell transcriptome were then tokenized based on a total vocabulary of 25,424 protein-coding or miRNA genes detected within Geneformer-30M. The token dictionary mapping each token ID to special tokens (pad and mask) or Ensembl IDs for each gene is included within the repository as a pickle file (token_dictionary.pkl).
 
 ### Data Fields
 
@@ -77,11 +79,11 @@ Mapping the gene regulatory networks that drive disease progression enables scre
 
 #### Initial Data Collection and Normalization
 
-Source data included 29.9 million (29,900,531) human single cell transcriptomes from a broad range of tissues from 561 publicly available datasets from original studies cited in the Extended Methods of Theodoris et al. 2022. Datasets were filtered to retain cells with total read counts within three standard deviations of the mean within that dataset and mitochondrial reads within three standard deviations of the mean within that dataset. Ensembl-annotated protein-coding and miRNA genes were used for downstream analysis. Cells with less than seven detected Ensembl-annotated protein-coding or miRNA genes were excluded as the 15% masking used for the pretraining learning objective would not reliably mask a gene in cells with fewer detected genes. Ultimately, 27.4 million (27,406,217) cells passed the defined quality filters. Cells were then represented as rank value encodings as discussed above in [Data Instances](#data-instances).
+Source data included 29.9 million (29,900,531) human single cell transcriptomes from a broad range of tissues from 561 publicly available datasets from original studies cited in the Methods of Theodoris et al, Nature 2023. Datasets were filtered to retain cells with total read counts within three standard deviations of the mean within that dataset and mitochondrial reads within three standard deviations of the mean within that dataset. Ensembl-annotated protein-coding and miRNA genes were used for downstream analysis. Cells with less than seven detected Ensembl-annotated protein-coding or miRNA genes were excluded as the 15% masking used for the pretraining learning objective would not reliably mask a gene in cells with fewer detected genes. Ultimately, 27.4 million (27,406,217) cells passed the defined quality filters. Cells were then represented as rank value encodings as discussed above in [Data Instances](#data-instances).
 
 #### Who are the source data producers?
 
-Publicly available datasets containing raw counts were collected from National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO), NCBI Sequence Read Archive (SRA), Human Cell Atlas, European Molecular Biology Laboratory-European Bioinformatics Institute (EMBL-EBI) Single Cell Expression Atlas, Broad Institute Single Cell Portal, Brotman Baty Institute (BBI)-Allen Single Cell Atlases, Tumor Immune Single-cell Hub (TISCH) (excluding malignant cells), Panglao Database, 10x Genomics, University of California, Santa Cruz Cell Browser, European Genome-phenome Archive, Synapse, Riken, Zenodo, National Institutes of Health (NIH) Figshare Archive, NCBI dbGap, Refine.bio, China National GeneBank Sequence Archive, Mendeley Data, and individual communication with authors of the original studies as cited in the Extended Methods of Theodoris et al. 2022.
+Publicly available datasets containing raw counts were collected from National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO), NCBI Sequence Read Archive (SRA), Human Cell Atlas, European Molecular Biology Laboratory-European Bioinformatics Institute (EMBL-EBI) Single Cell Expression Atlas, Broad Institute Single Cell Portal, Brotman Baty Institute (BBI)-Allen Single Cell Atlases, Tumor Immune Single-cell Hub (TISCH) (excluding malignant cells), Panglao Database, 10x Genomics, University of California, Santa Cruz Cell Browser, European Genome-phenome Archive, Synapse, Riken, Zenodo, National Institutes of Health (NIH) Figshare Archive, NCBI dbGap, Refine.bio, China National GeneBank Sequence Archive, Mendeley Data, and individual communication with authors of the original studies as cited in the Methods of Theodoris et al, Nature 2023.
 
 ### Annotations
 
@@ -101,7 +103,7 @@ There is no personal or sensitive information included in the dataset. The datas
 
 ### Social Impact of Dataset
 
-Genecorpus-30M enabled the large-scale pretraining of [Geneformer](https://huggingface.co/ctheodoris/Geneformer), a pretrained transformer model that enables context-aware predictions in settings with limited data in network biology. Within our publication, we demonstrated that during pretraining, Geneformer gained a fundamental understanding of network dynamics, encoding network hierarchy in the model’s attention weights in a completely self-supervised manner. Fine-tuning Geneformer towards a diverse panel of downstream tasks relevant to chromatin and network dynamics using limited task-specific data demonstrated that Geneformer consistently boosted predictive accuracy. Applied to disease modeling with limited patient data, Geneformer identified candidate therapeutic targets for cardiomyopathy. Overall, Geneformer represents an invaluable pretrained model from which fine-tuning towards a broad range of downstream applications can be pursued to accelerate discovery of key network regulators and candidate therapeutic targets.
+Genecorpus-30M enabled the large-scale pretraining of [Geneformer](https://huggingface.co/ctheodoris/Geneformer), a foundation model that enables context-aware predictions in settings with limited data in network biology. Within our publication, we demonstrated that during pretraining, Geneformer gained a fundamental understanding of network dynamics, encoding network hierarchy in the model’s attention weights in a completely self-supervised manner. Fine-tuning Geneformer towards a diverse panel of downstream tasks relevant to chromatin and network dynamics using limited task-specific data demonstrated that Geneformer consistently boosted predictive accuracy. Applied to disease modeling with limited patient data, Geneformer identified candidate therapeutic targets for cardiomyopathy. Overall, Geneformer represents a pretrained foundation model from which fine-tuning towards a broad range of downstream applications can be pursued to accelerate discovery of key network regulators and candidate therapeutic targets.
 
 ### Discussion of Biases
 
@@ -117,11 +119,12 @@ Genecorpus-30M was intended to be used for self-supervised pretraining. To achie
 
 Christina Theodoris, MD, PhD
 
-<!--- ### Licensing Information
+### Citation Information
 
-[More Information Needed]
+Theodoris CV*, Xiao L, Chopra A, Chaffin MD, Al Sayed ZR, Hill MC, Mantineo H, Brydon EM, Zeng Z, Liu XS, Ellinor PT*. Transfer learning enables predictions in network biology. Nature. 2023 May 31; Epub ahead of print. 
+(*co-corresponding authors)
 
-### Citation Information
+<!--- ### Licensing Information
 
 [More Information Needed]